Commodity video intelligent generation method based on multi-modal analysis and dynamic narrative architecture

Abstract

The present application relates to the technical field of video generation, in particular to a commodity video intelligent generation method based on multi-modal analysis and dynamic narrative architecture, inputting commodity original data and user behavior logs, outputting a dynamic selling point weight vector through a selling point value evaluation engine; inputting the dynamic selling point weight vector output by S1 into a narrative flow generator, executing: forming a narrative backbone chain according to the top K core selling points of the weight vector; generating a branch enhanced narrative flow based on user historical preference data insertion emotional reinforcement nodes; generating a multi-version narrative flow instruction set in combination with real-time network bandwidth data; generating an initial video sequence according to the instruction set to search a preset video segment library; detecting the semantic fault area between adjacent segments to generate compensation animation parameters; injecting the compensation animation parameters to output a continuous narrative video stream. The present application reduces the stuttering and frame skipping phenomenon caused by changes in narrative structure or different sources of video segments, and improves the overall viewing experience of users.

Description

Technical Field

[0001] This invention relates to the field of video generation technology, and in particular to a method for intelligent generation of product videos based on multimodal parsing and dynamic narrative architecture. Background Technology

[0002] With the rapid development of e-commerce and short video platforms, product videos have become an important way to showcase products for online sales. However, existing product video generation technologies generally have limitations.

[0003] Most product videos use a template-based structure, showcasing product features in a fixed shot order. This fails to personalize the content to cater to different users' interests, resulting in low relevance between video content and user interests, and inefficient information delivery. Traditional product videos often rely on static scripts or a pre-set order of selling points to construct the narrative chain, neglecting to incorporate real-time user behavior data or the complexity of product attributes for flexible editing. This leads to some key selling points that users care about not being presented first, impacting video retention and conversion rates.

[0004] Existing video splicing methods often suffer from visual jumps and abrupt rhythmic breaks when assembling multiple product video clips due to inconsistencies in content style, emotional tone, or pacing. Some technologies attempt to alleviate this problem using simple transition templates, but lack transition strategies based on semantic judgment and adaptive motion trajectory, resulting in limited compensation effects. Summary of the Invention

[0005] This invention provides a method for intelligently generating product videos based on multimodal parsing and dynamic narrative architecture. It integrates product image and text semantics, user behavior preferences, and network resource conditions to achieve a dynamically adjustable, clearly structured, emotionally connected, and continuously natural video generation method.

[0006] A method for intelligently generating product videos based on multimodal parsing and dynamic narrative architecture includes the following steps:

[0007] S1, Dynamic Selling Point Value Assessment: Input raw product data and user behavior logs, and the selling point value assessment engine outputs a dynamic selling point weight vector, including:

[0008] Parse product images and text descriptions to generate basic functional tag sets;

[0009] An attention distribution matrix is ​​generated by integrating user dwell time and click heatmaps.

[0010] Calculate the real-time value weight of each selling point based on behavior logs, and generate a dynamic selling point weight vector;

[0011] S2, Attention-Driven Narrative Flow Generation: Input the dynamic selling point weight vector output from S1 into the narrative flow generator and execute:

[0012] S21, based on the weight vector, extract the top K core selling points to form the main narrative chain;

[0013] S22, based on user historical preference data, insert emotion enhancement nodes to generate branches and enhance the narrative flow;

[0014] S23, combine real-time network bandwidth data to generate a multi-version narrative stream instruction set;

[0015] S3, fault-compensated video synthesis: The narrative stream instruction set generated in S2 is input into the video synthesizer, and an initial video sequence is generated by searching a preset video segment library according to the instruction set; semantic fault regions between adjacent segments are detected to generate compensation animation parameters; compensation animation parameters are injected to output a continuous narrative video stream.

[0016] Optionally, the step S1 of parsing the product image and text description to generate a basic functional tag set specifically includes:

[0017] The text and image descriptions of products are analyzed, including structured parameter information and natural language description text. Named entity recognition technology is used to extract key entity information of the products, including materials, efficacy, and applicable scenarios, and their corresponding semantic categories are labeled. Dependency parsing is performed on the text to identify the semantic dependencies between functional words and their modifiers. Then, a function-attribute association tree is constructed to establish semantic binding between each function and its specific attributes. Through semantic normalization, synonym merging, and context consistency verification, the extracted results are organized into a standardized set of basic functional tags.

[0018] Optionally, the construction of the attention distribution matrix in S1 includes:

[0019] Based on user behavior log data, a two-dimensional matrix reflecting the distribution of user attention is constructed. Dwell time sequence and click heatmap data are extracted from the user's interaction behavior during browsing product pages. The page display area is divided into multiple grid units, and the click frequency and gaze dwell time in each grid unit are counted. By integrating two sources of attention, click behavior and time behavior, the attention density value of each grid unit is calculated to form a spatial-temporal two-dimensional attention distribution matrix, which is used to characterize the degree of user attention in different areas of the product.

[0020] Optionally, the calculation of the real-time value weight in S1 specifically includes:

[0021] The system acquires a set of basic function tags and an attention distribution matrix, and associates the tag information with user attention behavior. Through front-end tracking and interface element recognition, it obtains the display area location of each basic function tag on the page and establishes a mapping relationship between tags and areas. It statistically analyzes user click behavior and attention distribution within the area corresponding to each tag, comprehensively calculates the real-time value weight of the tag, uses it to measure the user's current level of attention to the selling point, and uses it as a driving factor for narrative flow generation. Finally, it outputs a dynamic weight vector that includes all basic function tags.

[0022] Optionally, the construction of the S21 narrative backbone includes: receiving a dynamic selling point weight vector, sorting each selling point according to its value weight, calculating a truncation threshold for truncation based on the product's attribute complexity, extracting the top few key selling points to construct the backbone narrative chain, and forming a linear narrative backbone chain of the product's core functions.

[0023] Optionally, S22 includes the insertion of emotion enhancement nodes, specifically including generating an emotion preference vector based on the user's historical behavior data, determining whether each narrative node in the main narrative chain meets the set emotion activation conditions, and if so, inserting an emotion enhancement node matching the user's preference style at the corresponding position to expand a branch-enhanced narrative flow including branch paths.

[0024] Optionally, the generation of the multi-version narrative flow instruction set in S23 includes: matching the corresponding visual effect level standard according to the real-time network bandwidth status of the current terminal device, traversing all narrative nodes in the branch-enhanced narrative flow, the branch-enhanced narrative flow including the main chain node and the branch node inserted by the emotional enhancement mechanism, assigning the corresponding display duration and shot effect level to each narrative node, generating structured rendering instructions, and finally outputting multiple multi-version narrative flow instruction sets adapted to different bandwidth conditions, with each narrative node generating a triplet: [node ID, display duration, effect level].

[0025] Optionally, the generation of the initial video sequence in S3 includes inputting a multi-version narrative flow instruction set into a video synthesizer, parsing the node identifier, display duration, and special effects level information included in each instruction, retrieving the corresponding video unit from a preset video segment library according to the node identifier, and assembling them according to the instruction order to construct an initial video sequence including multiple video units. The video unit carries its start and end times, visual effects markers, and semantic description information.

[0026] Optionally, S3 further includes semantic tomography detection and compensation, specifically including performing tomography detection on adjacent video units in the initial video sequence, comprehensively considering two dimensions: semantic similarity and temporal continuity, and calculating the coherence score between adjacent units; when the coherence score is lower than a set score threshold, it is determined that there is a tomography region, and compensation animation parameters are generated accordingly, including compensation type, compensation duration, and transition trajectory obtained based on optical flow estimation between adjacent frames.

[0027] Optionally, S3 includes inserting the compensation animation between the fault regions, constructing a transition frame sequence, and recombining it with the original video units to generate a complete video stream; at the same time, adjusting the duration and aligning the beat of the audio track to output a continuous narrative video stream with consistent rhythm and visual continuity.

[0028] The beneficial effects of this invention are:

[0029] This invention constructs a multimodal evaluation system comprised of image and text analysis, user behavior fusion, and tag region mapping. This system dynamically generates a weighted vector of selling points that reflects current user preferences. This vector responds in real-time to data such as user dwell time and click behavior, ensuring that the narrative unfolds in the order most aligned with user interests. Compared to traditional static sorting methods, this dynamic weighting mechanism significantly enhances the display priority of key selling points, resulting in a narrative-driven increase in conversion rates.

[0030] This invention introduces user emotional preference vectors into product video generation, constructing a branching narrative flow with emotional enhancement nodes. Under the condition of satisfying significant emotional preferences, the system automatically matches stylized nodes that align with the user's aesthetic inclinations and inserts them into the main narrative chain, forming a dual-track display structure. This structure effectively solves the problems of lack of personalized information and monotonous content expression in existing technologies, enhancing user resonance and immersion without interfering with the delivery of key information.

[0031] This invention addresses the issue of content jumps in video narratives by proposing a two-factor tomography detection model based on semantic similarity and temporal continuity. It also introduces a dynamic compensation animation generation mechanism. The system adaptively selects the compensation type (scaling, displacement, fading) based on the scoring results and utilizes optical flow estimation to generate natural transition trajectories, achieving smooth transitions between tomographic regions. Simultaneously, an audio-visual synchronization mechanism ensures visual and rhythmic consistency in the transition content. This significantly reduces stuttering and frame skipping caused by changes in narrative structure or different video clip sources, improving the overall user viewing experience. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0035] like Figure 1 As shown, the intelligent generation method for product videos based on multimodal parsing and dynamic narrative architecture includes the following steps:

[0036] S1, Dynamic Selling Point Value Assessment: Input raw product data and user behavior logs, and the selling point value assessment engine outputs a dynamic selling point weight vector, including:

[0037] Parse product images and text descriptions to generate basic functional tag sets;

[0038] An attention distribution matrix is ​​generated by integrating user dwell time and click heatmaps.

[0039] The real-time value weight of each selling point is calculated based on behavioral logs, and a dynamic selling point weight vector is generated. Specific solutions are shown in S11-S13:

[0040] S11, Basic Function Tag Set Generation: Parse the structured data and unstructured text in the product image and text description, and perform the following processing:

[0041] Named Entity Recognition (NER) technology is used to extract entity information such as material, efficacy, and applicable scenarios of products. Based on dependency parsing, a semantic dependency structure between functions and attributes is constructed to form a function-attribute association tree, outputting a standardized set of basic function tags. ;in, This represents the set of functional tags for the product. For the first Functional labels.

[0042] S12, Construction of Attention Distribution Matrix: Generating a spatial-temporal two-dimensional attention matrix by fusing real-time user behavior logs, the steps are as follows:

[0043] Extract the sequence of user dwell time on the current product page: , This represents the total length of the sequence of time a user spends on the current product page (i.e., the number of time segments spent on the page).

[0044] Capture the coordinate data of user clicks on the heatmap: , This represents the total number of click events (i.e., the number of clicks) that occurred on the current product page.

[0045] Divide the product display interface into The grid, each grid cell is Calculate its attention density value:

[0046] ;

[0047] in, These are the row and column indices of the grid, respectively. This represents the index of the p-th click point in the click heatmap, where 1 ≤ p ≤ k. This represents the index of the q-th dwell time segment, during which the gaze rests on a grid cell. Inside, This represents the set of indices for all dwell time segments, in the range 1 ≤ r ≤ m. For the indicator function, if point Falling on the grid If the value is positive, it is 1; otherwise, it is 0. Indicates falling on the grid The set of stay time period indexes in the middle, An adjustable attention fusion weighting factor is used to balance the contributions of click behavior and dwell time, ultimately resulting in the attention distribution matrix: .

[0048] S13, Real-time Value Weight Calculation: Based on Basic Function Tag Set Attention distribution matrix Calculate the real-time value weight for each tag:

[0049] Establish a mapping relationship between labels and screen areas: ;in, Indicates label The display area in the interface.

[0050] Calculate the weight value for each tag:

[0051] ;

[0052] in, Represents the coordinates of a single click point Number of clicks The total number of all clicks. This represents the sum of attention intensity within the labeled area. This represents a moderating factor that controls the weighting distribution between click behavior and attention hotspots. These represent the row and column indices of the grid matrix (used for traversing the global grid). The final output is the real-time value weight:

[0053] Each Indicates function label Real-time value weight, It is a dynamic weight vector calculated based on user attention behavior and product display area.

[0054] In addition, the specific operations for extracting core product information based on named entity recognition are as follows:

[0055] When processing product descriptions (images and text), the system inputs the text portion into a pre-trained named entity recognition model to identify key information related to the product's core features, primarily including:

[0056] Material-related entities: such as "pure cotton," "stainless steel," and "ABS plastic," representing the composition of the product;

[0057] Functional items: such as "heat insulation", "antibacterial", "shock absorption", "sun protection", etc., which characterize the functional uses of the product;

[0058] Applicable scenario-based entities: such as "outdoor cycling", "baby products", "business meetings", etc., reflecting the usage environment or people.

[0059] To achieve the above functions, the NER model adopts a structure based on BERT or BiLSTM-CRF, and is trained and adapted in combination with a product category sub-vocabulary, so that it can accurately extract target entities from multiple product descriptions. When the model outputs, it labels each identified entity with its category (such as [material], [efficacy], [scenario]).

[0060] Furthermore, after completing NER extraction, dependency parsing is performed on the product description text. The main goal is to uncover the semantic structure of "which specific attributes correspond to a certain function". The specific operations for constructing the "function-attribute" semantic dependency structure based on dependency parsing are as follows:

[0061] Dependency parsing tools (SpaCy, HanLP, or Stanza) are used to construct syntax trees for product description sentences, extracting syntactic dependencies (such as subject-verb, modifier, complement, etc.). Using identified "functional words" as anchors, their dependency modification chains are expanded outwards to extract associated attribute words or modifiers, for example:

[0062] Enter: "This water bottle features a double-walled stainless steel structure, providing excellent heat retention."

[0063] Functional anchor point: Thermal insulation

[0064] Dependency path resolution: "heat preservation" ← "effect" ← "possess", forming the semantic unit "heat preservation effect".

[0065] Attribute association: "double-layer stainless steel structure" and "thermal insulation" establish a structural relationship through syntactic dependency.

[0066] The semantic relationships described above can be abstracted into a "function-attribute association tree": nodes are functional keywords, and child nodes are their related attributes and modifiers.

[0067] After obtaining the "functional-attribute association tree", perform the following standardization process:

[0068] Functional backbone extraction: For each functional semantic tree, extract its root node as a candidate functional label;

[0069] Semantic normalization: Use word vector clustering or thesaurus (e.g., group "antibacterial", "antimicrobial", and "sterilizing" into the same function) for classification and merging;

[0070] Attribute completion and fusion: While retaining the core semantics of functional tags, composite tags are generated by combining their modifier attributes (such as "thermal insulation_double-layer stainless steel").

[0071] Multi-source consistency verification: If the product's text and image description also contains structured parameters (such as specifications and functional parameter tables), these are used to supplement or correct the generated labels.

[0072] Output format: The final output is a set of standardized, machine-readable function tags.

[0073] Create tags With screen area The mapping relationship between these elements involves locating each functional label displayed on the product page to its specific display area on the user interface. This allows for subsequent calculation of the label's attention and weight based on user click heatmaps and dwell behavior. The specific implementation process is as follows:

[0074] 1. Interface Element Positioning: On the front-end product details page, HTML structure is used to display product features, such as: tag keywords in the text and image description area, function items in the parameter form, and selling point modules represented by visual icons. These elements often have unique DOM node IDs or CSS class names, and their display area information on the screen can be captured through front-end monitoring code (such as JavaScript tracking scripts).

[0075] 2. Display Area Coordinate Extraction: When the product page finishes loading, call the browser's getBoundingClientRect() interface to obtain the position information of the element corresponding to each functional tag, including its top-left corner coordinates, width, and height, i.e.:

[0076] Horizontal start and end points (xmin, xmax);

[0077] Vertical start and end points (ymin, ymax);

[0078] This rectangle defines the label. In the display area on the screen .

[0079] 3. Matching tag content with functional tag set: After executing S11, a standardized functional tag set 𝑇 has been obtained. At this time, it is necessary to compare the text tags displayed on the front-end page with this set: If the interface text matches a tag in the tag set (or matches successfully after merging synonyms), then a mapping relationship is established.

[0080] For a successfully matched tag ti, a binding is established with its corresponding rectangular region Ri.

[0081] 4. Construct the mapping function: This ultimately forms the mapping function from labels to regions. This means that the system can find the display area of ​​any function label in the interface and use it to calculate behavioral indicators such as user click density and dwell time in that area.

[0082] S2, Attention-Driven Narrative Flow Generation: Input the dynamic selling point weight vector output from S1 into the narrative flow generator and execute:

[0083] S21, based on the weight vector, extract the top K core selling points to form the main narrative chain;

[0084] S22, based on user historical preference data, insert emotion enhancement nodes to generate branches and enhance the narrative flow;

[0085] S23, which combines real-time network bandwidth data to generate a multi-version narrative flow instruction set.

[0086] S2 specifically includes the following steps:

[0087] S21, The narrative backbone is constructed based on a dynamic selling point weight vector: Sort by weight value from high to low to generate a priority queue for selling points:

[0088] ;

[0089] The backbone truncation threshold is dynamically calculated based on the complexity of product attributes.

[0090] Where, attribute complexity = number of functional tags User cognitive difficulty level This represents the floor function. Before extraction... The tags with the highest weights form the main narrative chain:

[0091] .

[0092] S22, Emotional Enhancement Node Insertion: Based on user historical preference data, enhance the expressiveness of the backbone chain:

[0093] First, analyze the user's historical behavior data to generate a sentiment preference vector: , This represents the user's preference for tech-themed content, with a value ranging from [0, 1]. This represents the user's preference value for luxury-style content, ranging from [0, 1]. This represents the user's preference value for minimalist content, with a value range of [0,1].

[0094] For each node in the backbone , This indicates the first [number] in the main narrative chain after sorting. Each functional label determines the emotional activation condition:

[0095] If the following conditions are met: ;

[0096] Then insert an emotion enhancement node at the corresponding position: ;

[0097] At the same time, the corresponding branch path is generated: ;

[0098] The final output is a branching narrative flow with emotional subplots. : Branch path .

[0099] This represents the maximum value among the three style preference values. This indicates the style category (tech / luxury / minimalist) that gives the highest emotional preference value. This indicates the inserted sentiment enhancement node. The emotional activation threshold is used to determine whether to insert an emotional node, and its value is 0.7.

[0100] Emotional activation threshold There needs to be a threshold for emotional triggering to ensure "preference salience". The purpose of the emotional activation mechanism is to insert the corresponding emotional reinforcement node only when the user has a clear emotional tendency towards a certain type. If the threshold is too low, it is easy for users without significant preferences to frequently trigger branches, which will interfere with the expression of the main narrative chain and destroy the focus of information transmission.

[0101] The maximum value in the normalized vector ≥ 0.7 is considered as the criterion for "preference dominance".

[0102] Experimental Verification: User Behavior Data Analysis Based on A / B Testing

[0103] In product prototype testing, A / B group experiments were conducted on multiple emotional activation thresholds (0.5, 0.6, 0.7, 0.8) to obtain threshold-emotional trigger control data:

[0104] Table 1 Threshold-Emotional Trigger Comparison Table

[0105] threshold θ Emotional Branch Trigger Rate User retention rate Branch bounce rate (interference perception) 0.5 High (>60%) middle High (>35%) 0.6 middle middle middle 0.7 Moderate (≈35%) High (↑11%) Low (<15%) 0.8 Low (<20%) middle Low

[0106] The results show that setting θ=0.7 can achieve a moderate effect of personalized narrative without excessively interfering with the main chain, which is the best balance between user acceptance and information completeness.

[0107] S23, Multi-version instruction set compilation: based on real-time network bandwidth data Compile multiple versions of effects instruction sets for each node in the narrative flow, and define the mapping relationship between effect levels and bandwidth:

[0108] Table 2. Mapping Relationship between Special Effects Level and Bandwidth

[0109] Bandwidth range (Mbps) Special effects level Lens type B≥20 L3 3D particle effects 10≤B<20 L2 Two-dimensional dynamic graphics B<10 L1 Static image and text translation

[0110] Branch traversal enhances narrative flow Generate a triple for each narrative node: [node ID, display duration, effect level];

[0111] Output the complete multi-version narrative flow instruction set: ; among them, each Corresponding special effects level The video rendering instruction set below.

[0112] S3, tomography-compensated video synthesis: The narrative stream instruction set generated in S2 is input into the video synthesizer. An initial video sequence is generated by retrieving a pre-defined video segment library based on the instruction set. Semantic tomography regions between adjacent segments are detected, and compensation animation parameters are generated. The compensation animation parameters are then injected, and a continuous narrative video stream is output. Specifically, this includes S31-S33:

[0113] S31, Initial Video Sequence Generation: Narrative Flow Instruction Set Input a video synthesizer and perform the following operations:

[0114] Parse each instruction triple: [node ID, display duration, effect level]; retrieve the corresponding atomic video unit from the video clip library based on the node ID; assemble according to the instruction order to generate the initial video sequence: ; among them, each Include the following metadata items:

[0115] [Start and end timestamps, special effects markers] ; Represents the first video sequence in the initial video sequence. Atomized video unit, For video unit The semantic descriptor vector.

[0116] S32, Semantic Tomography Detection and Compensation: Tomography detection is performed on adjacent units in the initial video sequence, as follows:

[0117] For each pair of adjacent units Calculate semantic coherence score:

[0118] If rating If a fault is detected, a compensation animation needs to be inserted, and compensation animation parameters need to be generated. Represents video unit End timestamp, Represents video unit The start timestamp, The maximum allowable time interval is used to normalize time series differences. This is the weighting coefficient for semantic similarity and temporal consistency in the coherence score. For the first Score the semantic coherence of adjacent video units. It is the threshold for determining the fault in the continuity score, with a value of 0.6 to 0.7.

[0119] S33, Continuous Video Stream Reconstruction: Based on the compensation animation parameters, the fault regions are repaired and the final video is output: Transition frame sequences are inserted at each fault location. ,Constructing and reconstructing the video:

[0120] Among them, the symbol This function represents the splicing operation of video clips; synchronizes the audio track with the animation compensation time length, aligns the audio and video as a whole; and outputs the final continuous narrative video stream. Indicates the sequence of compensated animation frames, inserted at between, To inject a complete, continuous narrative video stream after compensation, Narrative Flow Instruction Set The total number of nodes includes the total number of atomic video units corresponding to the main chain nodes and the emotion branch nodes.

[0121] Select the compensation animation type:

[0122] Based on fault scoring Given a numerical range, a compensation type selector is invoked to automatically determine which type of compensation animation to use, based on the following rules:

[0123] If the score is below a certain first threshold of 0.4 ( ), using scaling compensation, is suitable for situations with drastic semantic jumps and abrupt scene transitions;

[0124] If the score is in the middle range ( 6): Displacement-type compensation is used, which is suitable for narrative jumps with slight misalignment;

[0125] If the score is close to but does not reach the threshold ( ): Gradual compensation is selected for smooth transitions of minor semantic changes.

[0126] The duration of the compensation animation is dynamically calculated based on the fault score. The lower the score, the more severe the fault, and the longer the corresponding compensation animation duration. An "inverse proportional enhancement" method is adopted, that is, the lower the score, the longer the compensation time, to ensure that the screen transition is natural enough.

[0127] Finally, optical flow estimation is used to obtain the pixel motion trajectory between keyframes of two consecutive video units. The specific method is as follows:

[0128] extract The last frame and In the first frame, the dense optical flow algorithm is used to calculate the motion vector field between the two frames; the principal direction, velocity and structural changes are extracted from the motion vector as trajectory parameters for compensating the animation; this trajectory is used to guide frame interpolation or special effects rendering to achieve a seamless transition between the two video segments.

[0129] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0130] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent generation of product videos based on multimodal parsing and dynamic narrative architecture, characterized in that, Includes the following steps: S1, Dynamic Selling Point Value Assessment: Input raw product data and user behavior logs, and the selling point value assessment engine outputs a dynamic selling point weight vector, including: Parse product images and text descriptions to generate basic functional tag sets; An attention distribution matrix is ​​generated by integrating user dwell time and click heatmaps. Calculate the real-time value weight of each selling point based on user behavior logs, and generate a dynamic selling point weight vector; The process of parsing product image and text descriptions to generate a basic functional tag set specifically includes: The text and image descriptions of products are analyzed, including structured parameter information and natural language description text. Named entity recognition technology is used to extract key entity information of products, including materials, efficacy, and applicable scenarios, and their corresponding semantic categories are labeled. Dependency parsing is performed on the text to identify the semantic dependencies between functional words and their modifiers. Then, a function-attribute association tree is constructed to establish semantic binding between each function and its specific attributes. Through semantic normalization, synonym merging, and context consistency verification, the extracted results are organized into a standardized set of basic functional tags. The calculation of the real-time value weight specifically includes: The system acquires a set of basic function tags and an attention distribution matrix, and associates the tag information with user attention behavior. Through front-end tracking and interface element recognition, it obtains the display area position of each basic function tag on the page and establishes a mapping relationship between tags and areas. It also counts user click behavior and attention distribution within the area corresponding to each tag, calculates the real-time value weight of the tag, measures the user's current level of attention to the selling point, and uses it as a driving factor for narrative flow generation. Finally, it outputs a dynamic weight vector that includes all basic function tags. S2, Attention-Driven Narrative Flow Generation: Input the dynamic selling point weight vector output from S1 into the narrative flow generator and execute: S21, based on the weight vector, extract the top K core selling points to form the main narrative chain; S22, based on user historical preference data, insert emotion enhancement nodes to generate branches and enhance the narrative flow; S23, combine real-time network bandwidth data to generate a multi-version narrative stream instruction set; S3, fault-compensated video synthesis: The narrative stream instruction set generated in S2 is input into the video synthesizer, and an initial video sequence is generated by searching a preset video segment library according to the instruction set; semantic fault regions between adjacent segments are detected and compensation animation parameters are generated; the compensation animation parameters are injected and a continuous narrative video stream is output.

2. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 1, characterized in that, The construction of the attention distribution matrix in S1 includes: Based on user behavior log data, a two-dimensional matrix reflecting the distribution of user attention is constructed. Dwell time sequence and click heatmap data are extracted from the user's interaction behavior during browsing product pages. The page display area is divided into multiple grid units, and the click frequency and gaze dwell time in each grid unit are counted. By integrating two sources of attention, click behavior and time behavior, the attention density value of each grid unit is calculated to form a spatial-temporal two-dimensional attention distribution matrix, which is used to characterize the degree of user attention in different areas of the product.

3. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 1, characterized in that, The construction of the S21 narrative backbone includes: receiving a dynamic selling point weight vector, sorting each selling point according to its value weight, calculating a truncation threshold for truncation based on the product's attribute complexity, extracting the top few key selling points to construct the backbone narrative chain, and forming a linear narrative backbone chain of the product's core functions.

4. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 3, characterized in that, S22 includes the insertion of emotional enhancement nodes, specifically including generating an emotional preference vector based on the user's historical behavior data, determining whether each narrative node in the main narrative chain meets the set emotional activation conditions, and if so, inserting an emotional enhancement node that matches the user's preferred style at the corresponding position, thus expanding into a branch-enhanced narrative flow that includes branch paths.

5. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 4, characterized in that, The generation of the multi-version narrative flow instruction set in S23 includes: matching the corresponding visual effect level standard according to the real-time network bandwidth status of the current terminal device, traversing all narrative nodes in the branch-enhanced narrative flow, the branch-enhanced narrative flow includes the main chain node and the branch node inserted by the emotional enhancement mechanism, assigning the corresponding display duration and shot effect level to each narrative node, generating structured rendering instructions, and finally outputting multiple multi-version narrative flow instruction sets adapted to different bandwidth conditions, with each narrative node generating a triplet: [node ID, display duration, effect level].

6. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 5, characterized in that, The generation of the initial video sequence in S3 includes inputting a multi-version narrative flow instruction set into a video synthesizer, parsing the node identifier, display duration, and special effects level information included in each instruction, retrieving the corresponding video unit from a preset video segment library according to the node identifier, and assembling them according to the instruction order to construct an initial video sequence including multiple video units. The video unit carries its start and end time, visual effects markers, and semantic description information.

7. The intelligent product video generation method based on multimodal parsing and dynamic narrative architecture according to claim 6, characterized in that, The S3 further includes semantic tomography detection and compensation, specifically including performing tomography detection on adjacent video units in the initial video sequence, comprehensively considering two dimensions of semantic similarity and temporal continuity, and calculating the coherence score between adjacent units; when the coherence score is lower than a set score threshold, it is determined that there is a semantic tomography region, and compensation animation parameters are generated accordingly, including compensation type, compensation duration and transition trajectory obtained based on optical flow estimation between adjacent frames.

8. The intelligent generation method for product videos based on multimodal parsing and dynamic narrative architecture according to claim 7, characterized in that, The compensation animation parameters are injected between semantic gap regions to construct a transition frame sequence, which is then recombined with the original video units to generate a complete video stream. At the same time, the audio track is adjusted in duration and aligned with the beat to output a continuous narrative video stream with consistent rhythm and visual continuity.

Images